International Studies Quarterly (2013) 57, 52–64
Violence and Ethnic Segregation: A Computational Model Applied
to Baghdad1
Nils B. Weidmann
University of Konstanz
and
Idean Salehyan
University of North Texas
The implementation of the United States military surge in Iraq coincided with a significant reduction in ethnic violence.
Two explanations have been proposed for this result: The first is that the troop surge worked by increasing counterinsurgent capacity, whereas the second argument is that ethnic unmixing and the establishment of relatively homogenous
enclaves were responsible for declining violence in Baghdad through reducing contact. We address this question using
an agent-based model that is built on GIS-coded data on violence and ethnic composition in Baghdad. While we cannot
fully resolve the debate about the effectiveness of the surge, our model shows that patterns of violence and segregation
in Baghdad are consistent with a simple mechanism of ethnically motivated attacks and subsequent migration. Our modeling exercise also informs current debates about the effectiveness of counterinsurgency operations. We implement a simple policing mechanism in our model and show that even small levels of policing can dramatically mitigate subsequent
levels of violence. However, our results also show that the timing of these efforts is crucial; early responses to ethnic
violence are highly effective, but quickly lose impact as their implementation is delayed.
Ethnic conflicts often baffle observers as groups that had
once lived together in relative peace commit horrendous
acts of violence against one another. In places such as
the former Yugoslavia, Rwanda, and Burundi, people
co-existed in mixed neighborhoods for years before mass
violence lead to a hardening of ethnic identities and the
intensification of ethnic hatred (see Fearon and Laitin
2003). Ethnic violence also escalated after the 2003 United States-led invasion of Iraq, as the dominant mode of
violence shifted from attacks against an occupying force
to sectarian killings by Sunnis, Shias, and Kurds. The
bombing of the Shia al-Askari Mosque in Samarra in
2006 precipitated a spiral of violence that, by 2007,
looked to be out of control, leading many in the United
States to call for a complete withdrawal of troops. SunniShia violence in Baghdad was especially pronounced as
the feeble government and inadequate United States
forces were unable to prevent frequent attacks by the
Mahdi Army, Al-Qaeda in Iraq, and several other militias.
At the same time, the nature of Baghdad changed from a
city where the two main ethno-religious groups resided in
1
Authors’ notes. Previous versions were presented at the 2010 Annual Meeting of the American Political Science Association, Washington DC, the 2009
Meeting of the Peace Science Society, and the New Faces in Political Methodology III conference, Penn State University, May 2010. We are grateful to the
Empirical Studies of Conflict project at Princeton University (especially Jacob N.
Shapiro and Joseph H. Felter) for access to the SIGACTS data, and to the
International Conflict Research group at ETH Zurich for providing the computational infrastructure. Nils Weidmann gratefully acknowledges support
from the Air Force Office of Scientific Research (grant #FA9550-09-1-0314),
and from the Alexander von Humboldt Foundation (Sofja Kovalevskaja
Award). This and prior versions of this paper benefited significantly from
comments by Navin Bapat, Andrew Enterline, Cullen Hendrix, Jacob Shapiro
and Richard Stoll, as well as three anonymous reviewers at ISQ. An online
appendix is available at http://nils.weidmann.ws/.
mixed neighborhoods to one with well-defined ethnic
neighborhoods, separated in some areas by a security wall
(Bright 2007).
Beginning in 2007, the United States adopted a new
strategy (often called the “surge”) in which it increased
the number of troops in Iraq, hoping to contain the level
of violence through a significant security presence. In
addition, the United States sought accommodation with
certain militias, changed counterinsurgency tactics, and
placed greater emphasis on institution building. The
training of Iraqi government forces also increased significantly. These measures preceded a significant decline in
violence in 2008 and 2009, particularly in Baghdad,
although the exact causes of this decline are debated
(Matthews 2008; Ricks 2009). These patterns of ethnic
violence and residential segregation lead to important
questions for the academic and policy communities. What
explains the spike of violence in Baghdad and the subsequent decline in the number of attacks? How do these
events relate to inter-ethnic violence and forced migration elsewhere? What are the most important measures
needed to prevent mass killings and ethnic segregation?
While we focus on Baghdad in this paper, our analysis
will hopefully shed light on the dynamics of ethnic conflict more generally. For scholars and policymakers, a
proper understanding of the Iraqi case–particularly the
effectiveness of different interventions–can help guide
future action and avoid potential pitfalls.
Given a single case, it is difficult to provide a definitive
answer as to why violence escalated and declined as it did
in Baghdad. However, we can draw some conclusions
about plausible combinations of factors that may have
produced the observed dynamics of conflict. Did
improved counterinsurgency work? Or, did ethnic
Weidmann, Nils B. and Idean Salehyan. (2013) Violence and Ethnic Segregation: A Computational Model Applied to Baghdad.
International Studies Quarterly, doi: 10.1111/isqu.12059
© 2013 International Studies Association
Nils B. Weidmann and Idean Salehyan
partition help reduce levels of violence? To address these
questions, we rely on a computational model, or computer representation, where we experiment with different
hypothetical scenarios informed by current theories of
ethnic violence.2
In particular, we are interested in how combinations of
parameters interact to produce insurgent violence as well
as ethnic settlement decisions. Rather than a completely
artificial world, we use geo-referenced data on ethnic settlement patterns and violent attacks in Baghdad to
inform our model. Although we cannot perfectly re-run
history, we can construct a model based on actual data
and compare the results of our artificial simulations with
general empirical patterns. While it is difficult to draw
firm conclusions about a single case, our models allow us
to assess theoretically interesting combinations of parameters and we present a novel approach to validating
the output of our simulations. Therefore, our aim is to
shed light on theoretical debates in the literature on
insurgency and counterinsurgency, and provide some
insights–albeit tentative–into the dynamics of violence in
Baghdad.
We create a computational model with very simple
assumptions about intergroup behavior and local rules of
interaction. We populate our artificial world with “Shia”
and “Sunni” agents, along with a small share of violent
individuals in each ethnic community who attack members of the other group. Agents are assigned to neighborhoods according to actual ethnic maps of Baghdad.
When attacks against one’s group occur, agents in that
neighborhood can respond by moving to other, safer
neighborhoods, or by retaliating with their own violence.
Thus, we can represent processes of violence as well as
migration and settlement patterns. We can then determine how close our computational model comes to
resembling actual data on violence and ethnic settlement
patterns in Baghdad in order to draw conclusions about
the parameter combinations at work. Even though we
cannot say for certain what “actually” occurred in Baghdad and which factors were most important in turning
the tide of the insurgency, our models can help to
inform the debate over the efficacy of counterinsurgency
operations.
In addition to our Sunni and Shia agents, we allow for
a state policing mechanism to punish insurgents. According to many scholars, a strong central state is needed to
enforce order, punish violent individuals, and maintain
peace among ethnic groups (Tilly 1978; Muller and Seligson 1987; Posen 1993; Fearon and Laitin 2003). Thus, we
introduce a state actor—a counterinsurgent force—who
can punish insurgents for their transgressions and
remove them from the simulation, thereby reducing discontent and fear among our agents. The ability to detect
and capture insurgents is a parameter that we manipulate
in our simulations. By introducing counterinsurgency
efforts as a variable, we are able to shed light on debates
about the efficacy of the “surge” and of increased state
capacity in limiting violence.
In the following sections, we discuss current theories of
interethnic peace and violence. Then, we introduce our
agent-based model and its application to the Baghdad
case. Rather than a completely artificial world, we ground
2
For other applications of computational modeling to ethnic conflict and
war, see Epstein (2002); Hammond and Axelrod (2006); Bennett (2008);
Cederman (2008); Geller and Alam (2010); and Lustick, Miodownik and
Eidelson (2004).
53
our model in the current theoretical literature and use
actual data from Baghdad to inform our modeling exercise. After examining the parameter combinations that
could plausibly reproduce patterns of violence in Baghdad, we turn to a counterfactual analysis. In particular,
we look at how the level of policing affects violence and
ethnic segregation, altering both the effectiveness of
counterinsurgency and the timing of its implementation.
The simulations suggest three main findings. First, even
without assuming that people have a preference for living
with co-ethnics, we show that violence can dramatically
increase ethnic segregation as civilians search for safety.
Second, we are able to show that ethnic segregation
places natural limits on the extent of violence; that is, as
communities become more perfectly segregated, violence
declines even in the absence of effective counterinsurgency operations. Third, we show that small increases
in state policing can significantly reduce the level of
violence, but that it is far more effective if implemented
early on in an insurgency. Our final section offers some
concluding observations.
Theories of Interethnic Peace and Conflict
In multiethnic societies, most of the time, the vast majority of ethnic groups do not come into conflict with others. While ethnic divisions are ubiquitous, organized
violence between communal groups is relatively rare.
Ethnic cleavages are clearly necessary for ethnic conflict,
but they are not a sufficient cause, even when historical
grievances run deep (Mueller 2000; Fearon and Laitin
2003; Collier and Hoeffler 2004; Wilkinson 2004). For
this reason, scholars in the “Hobbesian” tradition have
argued that diminished state capacity to control violence
among communal groups is an essential ingredient in
fostering conflict (Posen 1993; Herbst 2000; Rose 2000;
Fearon and Laitin 2003). Even when latent tensions
between groups exist, the vast majority of people would
never contemplate murdering their neighbors and
co-workers; rather, a small set of extremists are predisposed to violence and a strong state is needed to punish
these offenders (Mueller 2000; Lake 2002). However,
once state capacity erodes and extremists begin to attack
members of the out-group, violence can quickly spiral out
of control, leading ethnic identities to harden (Fearon
and Laitin 2003). As the ethnic cleavage becomes
dominant, even moderates will seek safety from attack by
turning to their co-ethnics for support. Some—particularly young males—may be inclined to join militant
groups for safety as well as to seek revenge (Urdal 2006).
Those who are not willing or able to fight may seek safety
in numbers by residing among their ethnic brethren.
This is precisely the pattern that was seen in Iraq. After
the 2003 invasion, there were too few United States
troops on the ground to prevent interethnic conflict
(Enterline and Greig 2007), and the nascent Iraqi government was too feeble to provide order on its own. Sunni
and Shia extremists began to compete for influence, and
their voices became louder than that of moderates. Many
observers pointed out that prior to the invasion, Baghdad’s neighborhoods were quite mixed and intermarriage
between Sunnis and Shias was common. Some saw intermarriage and ethnic mixing as a stabilizing force that
could prevent an all out civil war (Telhami 2005). But
with rising violence, one observer noted that “mixed couples who symbolize Iraq’s once famous tolerance are
increasingly entangled by hate” (Raghavan 2007). As
54
Violence and Ethnic Segregation
extremists attacked members of the other sectarian
group, moderates were forced to choose sides if only for
safety. Interethnic warfare reached new heights after the
bombing of a Shia Mosque in Samarra in 2006, which
caused widespread fear and anger among the Shia community. Many in Baghdad joined the various militias
operating there, while others moved to neighborhoods
populated by their co-ethnics or fled the city entirely.
Many of Baghdad’s neighborhoods were “ethnically
cleansed” as one or the other group was forced out of
the area (Agnew, Gillespie, Gonzalez and Min 2008).
Although normatively questionable, ethnic unmixing
may provide one explanation for the decline of interethnic violence. Some scholars argue the best way to secure
peace is to create a partition, or physical separation,
between ethnic groups (Kaufmann 1996; Sambanis 2000;
Chapman and Roeder 2007). According to this argument,
once identities have hardened through violence and
hatred runs deep, ethnic groups engaged in bloody conflicts will not be able to trust one another enough to coexist peacefully. The best way to secure peace, then, is to
separate groups through some form of partition. Partition
may entail the complete separation of the state into two
new entities. One could also argue that ethnic segregation within a state, or even a city, serves many similar
functions by reducing contact between ethnic groups and
establishing autonomous enclaves. Along these lines, studies of race riots in the United States have shown that
decreasing levels of urban racial segregation increases
contact, competition, and the frequency of riots (Olzak
et al. 1996). While ethnic segregation may not be desirable, or even stable in the long-term, it may serve as an
immediate fix to large-scale bloodshed.
These claims have been given renewed attention in the
context of the Iraq war. Joseph and O’Hanlon (2007:5)
argue for a “soft partition” of Iraq and claim that “physical separation boosts security, but keeping the communities cheek-by-jowl makes residents angry and resentful
(p. 5).” Even United States military planners agreed,
arguing that the Baghdad security wall, which separates
ethnic communities, “is one of the centerpieces of a new
strategy by coalition and Iraqi forces to break the cycle of
violence” (Wong and Cloud 2007). Researchers at the
University of California, Los Angeles, using satellite imagery of nighttime lights in Baghdad to determine settlement patterns, concluded that the military surge was
ineffective, but rather, ‘the diminished level of violence
in Iraq since the onset of the surge owes much to a
vicious process of interethnic cleansing” (Agnew et al.
2008:2295). In other words, Agnew et al. make a strong
case that ethnic segregation in Baghdad, as discerned
through satellite imagery, was in place prior to the troop
surge and may explain declining levels of violence.
Therefore, one hypothesis, which we turn to later, is
that ethnic segregation reduces ethnic violence. However,
as an alternative to the ethnic cleansing of neighborhoods and the creation of mutually-exclusive enclaves,
steps may be taken to strengthen the state and reestablish
control. If the lack of a robust deterrent capability and
inability to punish defectors led to ethnic violence in the
first place, then measures to bolster the central government may help to reverse violence. Many would agree
that—assuming that it is feasible—an impartial, competent state to resolve disputes and establish a monopoly of
legitimate force is normatively preferable to ethnic cleansing. State strength clearly means many things, including
the ability to provide public goods and execute adminis-
trative functions, but one important component of state
strength is the ability to punish criminals and violent
individuals (Hendrix 2010). If the state can detect and
defeat insurgents as well as provide civilians adequate levels of protection (Kalyvas 2006), conflict levels may come
down, internal displacement may subside, and over the
long-run, ethnic trust may be established. In his focus on
security institutions, Herbst (2004: 367) argues that in
order to reduce civil violence it is, “...essential to promote
institutions—notably the police and the military—that are
immediately responsible for order.” Fearon and Laitin
(2003) make a similar argument, and advocate increasing
the state’s capacity to govern its territory and root out
insurgents.
The troop surge in Iraq promised to accomplish many
of these functions. The United States sent more than
20,000 troops to Baghdad in 2007 and, along with Iraqi
forces, implemented Operation Fardh al-Qanoon (Operation Imposing Law). Baghdad was divided into several
zones and the United States military worked to clear each
area of hostile Sunni and Shia militias. In particular, military presence in residential areas was improved by setting
up “combat outposts” in different neighborhoods of
Baghdad (Londono 2007). The combat outposts were
used to ensure a quick response in cases of insurgent
attacks, but also to work more closely with the local population. United States forces may be seen as a “neutral”
third party enforcement mechanism, not tied to any
ethnic or religious faction, but with a mandate to enforce
order over the entire city. Simultaneously, the United
States engaged in efforts to strengthen the Iraqi state by
training troops and police officers while encouraging
broad participation in the government by all groups. The
stated goal was to secure Baghdad and provide the space
needed for ethno-sectarian groups to find an acceptable
accommodation and integrate into a strong, unified force
that could provide security for the country as a whole.
Many have argued that the troop surge in Iraq was a
success. United States and Iraqi casualties declined significantly after thousands of troops were deployed to Baghdad and elsewhere in 2007. In his report to Congress, the
United States Commander in Iraq, General David Petraeus highlighted improved counterinsurgent capacity, stating, “One reason for the decline in incidents is that
Coalition and Iraqi forces have dealt significant blows to
Al-Qaeda-Iraq... We have also disrupted Shia militia
extremists...”3 In the Wall Street Journal, United States
Senators John McCain and Joe Lieberman argued, “As
Al-Qaeda has been beaten back, violence across the
country has dropped dramatically... As the surge should
have taught us by now, troop numbers matter in Iraq”
(McCain and Lieberman 2008). Finally, in a speech in
2008, Lt. General Ray Odierno attributed the success of
the surge to, “attacking the enemy,” “establishing and
maintaining a presence in places that had long been
sanctuaries of Al-Qaeda,” and “going after Shia extremists.” He places special emphasis on the training of Iraqi
forces, stating that, “the surge of Coalition forces also
helped bring about a surge in Iraqi Security Force
capacity” (Odierno 2008). As these statements indicate,
increasing the number of United States troops, while
simultaneously improving the efficacy of the Iraqi
government, were seen as important components of the
counterinsurgency effort.
3
Report to Congress on the Situation in Iraq. General David H. Petraeus,
Commander, Multi-National Force-Iraq. September 10–11, 2007.
Nils B. Weidmann and Idean Salehyan
The troop surge in Iraq was clearly a multi-pronged
strategy that went beyond simply increasing troop levels.
However, one of the most important objectives of the
surge was to increase United States and Iraqi ability to
identify and eliminate insurgent forces. Part of this strategy included a change in tactics that de-emphasized heavy
firepower and focused on establishing ties to local communities (see Lyall and Wilson 2009). This change in
thinking on counterinsurgency operations and empowering locals to help with policing efforts were part of an
overall strategy to better identify and defeat violent
extremists. Building state capacity, moreover, is widely
viewed by scholars (discussed above) as a key component
of deterring and ending an insurgent campaign.
Therefore, there are two distinct arguments that have
been made with respect to declining violence in Baghdad. The first is that the troop surge worked by increasing counterinsurgent capacity. Increasing the number of
security forces in Baghdad, strengthening the Iraqi state,
and changing tactics to better identify insurgents all
helped to remove extremists from the population and
reduce levels of violence. The second argument is that
the troop surge was irrelevant. Instead, according to this
argument, ethnic unmixing and the establishment of relatively homogenous enclaves were responsible for declining violence in Baghdad through reducing contact.
Moreover, ethnically homogenous neighborhoods make
it more difficult for insurgents from the other group to
operate, since insurgents rely on networks of local support and it is relatively easier to identify individuals that
do not “belong” to the community. Unfortunately, the
Baghdad case is overdetermined as both of these processes occurred, and it is difficult to disentangle the relative strength of these claims with a single case. While we
cannot fully resolve this debate, we develop an agentbased model that takes into account levels of violence,
ethnic settlement patterns, and counterinsurgency effectiveness, to assess theoretically interesting combinations
of parameters and we validate our modeling exercise with
actual data taken from Baghdad.
In addition to looking at how ethnic settlement patterns may have influenced violence in Baghdad, we are
also able to assess counterfactual claims regarding state
capacity. In particular, we assess how effective augmenting counterinsurgent ability is in reducing violence
through experimental “treatments” in our simulation. For
any given level of counterinsurgent capacity, how much
violence will we see? We are also able to introduce
enhanced policing at various periods in the simulation to
address important questions regarding the timing of
counterinsurgent campaigns. Some, most notably United
States Army Chief of Staff Erik Shinseki, argued that a
strong troop presence should have been put in Iraq
immediately after the 2003 invasion in order to prevent
widespread violence. Along these lines, Enterline and
Greig (2007) find that a significant troop presence early
in a counterinsurgency campaign, the “Shinseki plan,” is
far more effective than troops added at a later date. By
varying the level and timing of state policing, we can see
how the dynamics of violence change in our models.
Model
A multitude of influences affected the dynamics of violence and segregation in Baghdad, and any attempt to
understand the micro-mechanisms behind them seems
difficult at best. Recognizing that we will not be able to
55
perfectly “re-run” Baghdad history, we aim to find out
whether the general interrelationships between violence
and migration we discussed above bear some resemblance
to the observed patterns. Similar to many other applications in the social sciences, this attempt needs to reflect
the micro/macro distinction made in Coleman’s “bathtub” (Coleman 1990): The mechanisms we postulate are
located at the micro-level, with insurgents deciding to
attack in certain neighborhoods, triggering migration of
civilians as a result. However, we can only observe the
result of this process at the macro-level—the uneven
distribution of violence across the city, and the stepwise
segregation of ethnic groups.
One way to fill the gap between micro-mechanisms and
macro-outcomes is by means of computational (or more
precisely, agent-based) modeling (ABM). Cederman
(2001: 16) defines ABM as “a computational methodology
that allows the analyst to create, analyze, and experiment
with, artificial worlds populated by agents that interact in
non-trivial ways and that constitute their own environment.” Still, the implementation of this methodology far
from standardized. First, ABMs differ widely as to the
complexity of the computational agents and their behavior. At the lower end of this complexity scale is
Schelling’s (1971) famous model of neighborhood
segregation, where the only feature of an agent is its
“color” and one simple rule—affinity for living with
agents of the same color—drives their behavior. Axelrod’s
(1984) computer simulations are also quite simple, with
agents playing the familiar Prisoners Dilemma game in
multi-player tournaments. These models are easy to
understand, but may be too basic to represent reality.
More recent examples such as Geller and Alam (2010) or
Cioffi-Revilla and Rouleau (2010) have grown far more
complex, featuring representations of entire countries
with political and environmental sub-systems. The intention of this approach is to create a more realistic representation of the world, but this comes at a cost. With a
model guided by dozens if not hundreds of parameters,
one quickly encounters what de Marchi (2005) calls the
“curse of dimensionality”: The complexity of the model
itself makes it difficult to understand the interplay of its
elements.
Along a second dimension, ABMs differ with respect to
the level to which they incorporate empirical data, if at
all. Computational models can serve as purely heuristic
devices, to illustrate how a theorized micro-process leads
to the emergence of macro-outcomes. Again, Schelling’s
(1971) model is an example here, but more recent ones
include Cederman’s (1997) model of war and state formation, Bhavnani’s (2006) norms model of mass participation in interethnic violence, or Siegel’s (2009) model
of collective action in social networks. Models that incorporate empirical data can do so by comparing a generated output to an observed pattern. For example,
Cederman (2003) shows how a simple model of interstate
war generates a power-law distribution of war sizes, similar
to what has been observed in reality. A similar approach
was taken by the authors of the “Artificial Anasazi” model
(Dean, Gumerman, Epstein, Axtell, Swedlund, Parker
and McCarroll, 2006), who managed to reproduce the
settlement pattern of an ancient culture in an agentbased model. Geller, Rizi and Łatek (2011) use actual
data from Afghanistan to validate a computer simulation
of the interaction between drug production and corruption. Despite some promising attempts, the empirical
evaluation of computational models is the exception
56
Violence and Ethnic Segregation
rather than the norm. As we will outline in the next paragraphs, our computational model adopts an intermediate
approach between simplicity and complexity by keeping
the number of moving parts within manageable ranges.
At the same time, we conduct a model fitting exercise by
letting empirical data guide the selection of plausible
parameter values. In the following paragraphs, we
describe our implementation of this model, and the next
section discusses the incorporation of empirical data.
The fundamental building blocks of our model is a
space with a set of locations that represent the neighborhoods of Baghdad, and the agents populating that space.
Agents can belong to one of two ethnic groups, Sunni or
Shia. In addition, Sunni and Shia agents are designated
to be either insurgents or civilians. The basic dynamics of
the model are as follows. Insurgents of one group
attempt attacks against civilians of the other group. The
success of an attack depends on the local ethnic makeup
of the location at which it is conducted. This generates
fear in the targeted population, which leads civilian
agents to consider migration to safer places in the city.
This in turn alters the ethnic configuration, which can
either increase or constrain the susceptibility for violence
in these locations. In the following, we first describe the
model structure (the agents and the model space) and
then turn to the dynamics of the model.
Model Space and Agent Population
The model space consists of N locations i that are populated by agents of two groups j, j ∈ {Sunni, Shia}. For a
given group j, we refer to the other group as ¬j. The initial total number of agents at each location is assigned at
the beginning of the simulation, and consists of both
insurgents I and civilians C. The proportion of insurgents
p is a model parameter that determines the size of the
insurgent population. Locations differ with respect to
their Sunni/Shia balance. As we will describe below, the
initial ethnic balance is taken from an ethnic map of
Baghdad. A certain percentage of agents at a location is
randomly chosen to be insurgents. This percentage is
equal across neighborhoods, and is a model parameter
that is fixed at the beginning of a model run. Throughout the simulation, we measure the ethnic composition
of a location at a given time simply as the proportion pi;j
of agents of group j at location i. Once the space has
been populated with insurgents and civilians of different
groups, the model proceeds in a series of time steps t as
described in the following paragraphs.
Violence
At the beginning of a time step t, each insurgent Ii;j
attempts to stage an attack at its current location i against
the members of the other group ¬j. Even though all
insurgents aim to attack, only few of these will be successfully carried out. We assume that the probability of success of an attack depends on the local ethnic
configuration of the location, as measured by the proportion of the insurgent’s co-ethnics in the respective location pi;j;t . Theoretically, this effect could take one of two
forms. On the one hand, interethnic violence could be
carried out as ethnic cleansing of small minorities of the
other group, for example by violent raids. If that is the
case, we expect the impact of the proportion of co-ethnics to be positive, implying that as neighborhoods
become increasingly homogenous, there is more pressure
for the minority group to leave. On the other hand,
attacks against a group could increase with its populations share as attackers try to cause maximum damage to
as many of these individuals as possible. A Shia bombing
of a marketplace in a Sunni neighborhood would be an
example of this type of attack, as it attempts to maximize
casualties in the other group. If most attacks take this
form, the effect of the proportion of co-ethnics would be
negative. These are competing claims that we can evaluate in our model. In addition to this local impact on
attack success, we want to allow for the possibility that the
insurgent attacks are unaffected by local conditions, and
are simply occurring with a constant rate of success across
all neighborhoods. In order to incorporate these competing alternatives, we let a logit equation determine the
probability of success:
pAttacki;j;t ¼ logit1 ða0 þ a1 pi;j;t Þ
ð1Þ
We then simply take a Bernoulli draw with the computed probability to determine whether an attack is successful. The intercept a0 and the coefficient a1 are model
parameters selected at the beginning of a simulation run,
which then drive all attack behavior during the subsequent time steps. Once insurgents have attempted to
attack, they relocate to a randomly selected location.
The logit model allows for a flexible specification of the
attack mechanisms we are interested in. First, as described
above, it makes it possible to determine whether a constant
attack probability is sufficient to explain the patterns we
observe in Baghdad, or whether the local ethnic mix does
indeed have an effect as we hypothesized. If the latter, we
should estimate a1 to be different from 0. Depending on
the sign of a1 , the effect can go in one of two directions.
First, if a1 > 0, a higher proportion of co-ethnics makes
attacks more likely to be successful. This is consistent with
the first mechanisms of ethnic violence we described
above, where violence is carried out to cleanse neighborhoods dominated by one group from members of the
other group. On the contrary, if a1 < 0, lower proportions
of co-ethnics will lead to more attacks. This is essentially
the car bomb violence mechanism, where the insurgent
attempts to maximize damage to the other group, but at
the same time avoids harm to his co-ethnics.
The result of the attack sub-procedure is for each
neighborhood, a number of attacks ai;j;t against a particular group. This in turn affects the civilian population as
we elaborate in the next paragraph.
Migration
Once insurgents have carried out their attacks, civilians
react to the threat and decide whether to relocate to
safer areas. We consider two potential influences of insurgent violence. First, attacks in a given neighborhood
directly influence the level of fear experienced by members of the victimized group. Here, it is the experienced
violence ai;j;t that affects an agent’s decision. Second, we
argue that violence can have effects beyond the borders
of a single neighborhood. Violent acts against one group
could instill fear in agents in other neighborhoods
nearby. We model this spillover of fear by including the
spatially weighted number of attacks as a determinant of
a civilian’s migration decision. Following common pracW
as the average numtice, we compute the spatial lag ai;j;t
ber of attacks against group j across all other
neighborhoods, weighted by the normalized inverse distance to neighborhood i. This spatial lag will take high
57
Nils B. Weidmann and Idean Salehyan
values if violence in proximate locations against group j is
high, but low values if this violence occurred in distant
neighborhoods. Again, we use a logit model to compute
the probability of migration for a civilian of group j at
location i, as given by the following equation:
W
pMigrationi;j;t ¼ logit1 ðb0 þ b1 ai;j;t þ b2 ai;j;t
Þ
Iraq (HIC).4 Using these neighborhoods as our basic
units of observation, we incorporate real data on the
level of violence and the ethnic composition of each
neighborhood.
ð2Þ
As for the attack model above, the migration model
allows us to separate the baseline migration from the one
that is (directly or indirectly) induced by ethnic violence.
The constant b0 and the coefficients b1 and b2 are model
parameters fixed throughout a simulation run. After comparing our model runs to actual data, if we observe b1
and b2 to be positive in empirically plausible runs, we
interpret this as evidence in support of our proposed violence-induced migration mechanism.
If an individual chooses to migrate, she does so by selecting a location that appears to be safer than the one she
is currently residing at. Having observed the insurgentgenerated violence in the current time step, she moves to a
randomly selected location z that experienced lower levels
of violence against her group, i.e. with az;j;t \ ai;j;t . Importantly, our models make no assumptions about
ethnophilia, or an inherent desire to live with one’s coethnics, in contrast to Schelling’s (1971) models. Migration decisions are driven by safety concerns, not ethnic
attachments.
In sum, our model relies on a set of six parameters.
First, we have the proportion of insurgents p in the
model. Next, the attack model is guided by the constant
a0 and the coefficient a1 , and the migration model by b0 ,
b1 and b2 . The purpose of our empirical analysis will be
to find out which parameter settings can explain the variation in violence and segregation we observe in Baghdad,
and in particular, if segregation alone is sufficient to
explain the decrease in violence. Later, in our counterfactual experiments, we will add a simple punishment mechanism to the model to evaluate a different way of
violence reduction and its effectiveness.
Violence
Data on incidents of sectarian violence was obtained from
the Iraq SIGACTS database.5 From all the events we
selected those that correspond to sectarian violence by
excluding (i) attacks involving United States/Iraqi security
forces, and (ii) minor/non-deadly event categories such as
“looting”. The SIGACTS database in its present version
covers the period from April 2004 through February 2009,
but the information on violence initiator and target is available only in a comprehensive version of the database that
covers April 2006 through September 2007. SIGACTS contains data at the event level. Each line in the main table list
one event along with its spatial and temporal coordinates.
Using the spatial coordinates of an event, we matched
events to our units of observation to compute event counts
of violence at the neighborhood level.
Ethnicity
We obtained estimates about the ethnic distribution at
the neighborhood level from M. Izady at the Gulf 2000
project at Columbia University.6 The project provides
maps of the predominant ethnic group in a neighborhood for different periods (2003, 2006, early 2007, late
2007, 2009). The map shows areas of Baghdad either as
mixed, Sunni- or Shia-dominated.7 We geo-referenced the
maps and created a similar variable (Sunni/Shia/mixed)
at the neighborhood level. Unfortunately, the data does
not provide precise estimates of group proportions—that
is, we cannot say if a “dominated” neighborhood is 70%
or 80% populated by one group—but we can use the
maps to ascertain overall settlement patterns.
Period of Observation
Data
Similar to many existing approaches (Schelling 1971;
Epstein and Axtell 1996; Cederman 1997), our computational model as described above could be implemented
in an entirely artificial space that uses cells as spatial
units. However, since our intention is to apply the model
to a particular case (Baghdad), we create a model that
shares certain characteristic features of our case: First, we
create the model based on the real geography of the city,
using neighborhoods as the unit of observation. Second,
we use observed incidents of sectarian violence to get the
distribution of violence across neighborhoods. Third, the
ethnic composition of neighborhoods in the model is
based on high-resolution ethnic maps of Baghdad. This
section explains our data sources and how they were integrated in the model.
Unit of Observation
The model operates on the set of 85 Baghdad neighborhoods, which were geo-referenced from a map provided
by the UNs Humanitarian Information Center for
4
Of the original 89 units, 4 primarily industrial areas with few residential
homes were excluded.
The bombing of the Askari Mosque in Samarra in February 2006 is commonly taken as the most significant trigger of violence between Sunni and Shia groups in Iraq.
During that year and especially in the autumn, violence
levels in Baghdad increased steadily. This was part of the
reason for the United States to adopt a new military strategy, the surge. Part of the implementation of the surge
was the additional deployment of United States military
personnel, primarily to Baghdad, where most of the violence in Iraq occurred at that time. This deployment
started in late January 2007, increased rapidly in February
and March, and reached its peak in June 2007. Thus, the
Samarra attacks in February 2006 can mark the onset of
inter-ethnic violence in Iraq, until the implementation of
the surge directly aimed to stop the hostilities between
groups. Thus, we use data from the period between the
5
Joseph H. Felter and Jacob N. Shapiro, Princeton University, 2008; see
Berman, Shapiro and Felter (2011) for an application.
6
M. Izady, Ethnic-Religious Neighborhoods in Metropolitan Baghdad.
Available at http://gulf2000.columbia.edu/maps.shtml. (Accessed February 4,
2010)
7
The maps also show Christian populations in Baghdad. However, there
is no homogenous Christian neighborhood in Baghdad due to their small
population share (<5%), so we do not treat them as a separate group in our
model.
58
Violence and Ethnic Segregation
Ethnic Groups (2006)
Violence (4/2006−1/2007)
Ethnic Groups (early 2007)
Ethnic Groups (early 2007)
Violence (2/2007−9/2007)
Ethnic Groups (late 2007)
FIG 1. Empirical data used for seeding and validation of the model. Ethnic maps show Shia (grey), Sunni (black) and mixed neighborhoods
(striped). The level of violence by neighborhood is displayed in different grey shades (center map)
Samarra bombings and the onset of the surge for our
calibration of the model. Figure 1 illustrates the ethnicity
and violence data for our observation period.
Results
Our approach of incorporating empirical data into a
computational model shares the essential features of the
Bayesian approach to statistics, in which we start with a
theoretical prior and update based on the incorporation
of empirical data. Essentially, our approach is as follows:
1. We develop a simple model of violence and migration, with a set of parameters determining its behavior. This model was described above.
2. The model is run with parameter combinations
sampled randomly from a large parameter space H
assuming no prior knowledge of the distributions of
the individual parameters (our “noninformative
priors”).
3. We select those parameter combinations H0 H that
produce empirically plausible runs, i.e. those that
come closest to the empirical patterns in Baghdad.
4. We examine the distribution of the parameter values
in H0 (our “posterior distributions”). The examination of this posterior distribution of the model
parameters reveals if a parameter is necessary to generate empirically plausible model outcomes, and if
so, whether it has the expected sign.
Model Initialization and Dynamics
As described above, we conduct many different runs of
the model with different parameter vectors h, and evaluate which parameter settings come closest to the observed
patterns of violence and segregation in Baghdad, i.e. produce the best fit. Using the data we have available for our
observation period (ethnic distribution at the beginning
and the end of the period, and distribution of violence),
we initialize the model based on this data, and compute
its goodness of fit as the simulation unfolds. More precisely, we initialize the model run with the ethnic distribution at the beginning of this period. This is done by
populating each neighborhood with 100 agents, assigning
a 50/50 proportion of ethnicity in mixed neighborhoods
(as given by our ethnic maps), and a 70/30 proportion
in neighborhoods dominated by one group.8 Since we do
not have data on absolute population at the neighborhood level, this procedure omits differences in population totals across neighborhoods. Figure 2 shows a
screenshot of the model after initialization.
Once it has been initialized, the model is run for a certain number of time steps. As described in detail above,
each step consists of the insurgents carrying out their
attacks, and the civilians reacting to violence. All of these
actions are determined by the model parameters h set for
the current run. Empirically plausible runs are those that
approximate both the ethnic distribution at the end of
the observation period, and the spatial pattern of violence during this period. In other words, we want those
runs that minimize the deviation of the model from the
observed ethnic distribution at the end of the observation
period, and second, those that at the same time produce
a distribution of violence that is as close as possible to
the observed one. Thus, we compute two error measures
for the model. ee ðtÞ is the proportion of incorrectly predicted neighborhood types (Sunni, Shia, or mixed). We
count a neighborhood as a correct prediction if at time t
it is dominated ( 70%) by the same group as given on
the ethnic map, or alternatively, if no group reaches a
70% share in the simulation and the neighborhood is
“mixed” according to the map. The error ev ðtÞ between
the observed violence map and the simulated one at time
t is simply the mean squared difference across neighborhoods between the observed normalized violence score
and the normalized simulated number of attacks per
8
Results are robust to using alternative proportions for dominated neighborhoods; see online appendix.
Nils B. Weidmann and Idean Salehyan
FIG 2. Screenshot of the computational model, initialized using the
2006 ethnic map. The shading indicates the group distribution
(bright: Shia, dark: Sunni)
neighborhood that occurred until time t (remember that
we would like to approximate the cumulative distribution
of attacks). As described below, the error measures ee ðtÞ
and ev ðtÞ help us select those parameter combinations of
the model that come closest to the observed patterns in
Baghdad.
Parameter Estimates
First, we fit our model to the empirical data described
above. This exercise shows us whether the model parameter estimates we obtain are in line with our theoretical
reasoning. These parameter estimates help us conduct
simple counterfactual experiments with our model,
described in the next section.
Following our strategy described above, we run the
model on the empirical data, using 20,000 randomly
drawn parameter vectors. In principle, the coefficients in
the attack and migration logit models could be
unbounded. Practically, however, we only need to consider parameter ranges where the coefficients in the logit
models have a significant impact on the computed probability; for this reason, we restrict the sampling range for
the coefficients a1 , b1 and b2 to [10,10] and for the
intercepts a0 and b0 to [20,0]. For the experiments
reported below we use a fixed proportion of insurgents p
of 0.02, indicating that only a very small share of the population is disposed to violence.9 Each run is limited to
500 time steps.10 We now need to select those parameter
vectors h that maximize the fit with respect to both the
distribution of groups and violence across the city. More
precisely, a “plausible” parameter vector eventually (i)
produces a simulated ethnic map that approximates the
observed one, (ii) produces a spatial distribution of violence similar to the observed one, and (iii) does so at the
same time during the model run. Note that since we do
not calibrate the model with respect to time, the model
can reach a good fit at any time during the simulation.
In order to select parameter configurations based on
these three criteria, we record for each model run (i) the
time topt when ee ðtÞ is minimal: topt ¼ arg mint2½0;500 ee ðtÞ,
9
Our tests with alternative values of this parameter (0.01, 0.03) produce
similar results; see online appendix.
10
We conducted initial experiments showing that most of the model
dynamics occur within the first 100–200 time steps.
59
(ii) the value of ee ðtopt Þ, and (iii) the value of ev ðtopt Þ.
Remember that model runs start with an ethnic configuration as given by the ethnic map at the beginning of the
respective period, so ee ð0Þ is simply the prediction error
between this map and the one at the end of the period.11
Many parameter combinations yield topt ¼ 0; in other
words, the model never improves upon the ethnic map at
the beginning of the period. Moreover, a small topt could
indicate that the improvement in ee is due to random
fluctuation at the beginning of the model run. Therefore,
when selecting our plausible parameter vectors, we first
discard all runs with topt 5, and from this sample, select
those where both ee ðtopt Þ and ev ðtopt Þ are smaller than
their respective mean.
This results in a set H0 of 533 plausible parameter
vectors. This low number emphasizes the narrow margin within which the model captures the dynamics similar to Baghdad: Only about 3% of all parameter vectors
are likely to have generated our data. In other words,
the vast majority of parameter combinations can be
ruled out, leaving us with a more limited range of theoretically interesting models. The fit of the model is
fair but not overwhelming. In the pre-surge period, the
model reduces the prediction error ee for ethnicity
from an ee ð0Þ of 0.45 to a minimum of 0.38. The minimal error ev between the simulated and the observed
violence maps is 0.05. Still, given the lack of data for
many essential features in Baghdad, a perfect replication of migration and violence for each of the 85
neighborhoods is near impossible. Rather, our models
are able to capture the most important dynamics of
the empirical case: Increasing segregation along with
decreasing violence. Most importantly, we are able to
exclude many implausible parameter combinations and
can focus attention on the most theoretically interesting
estimates.
Proportion of Co-ethnics and Insurgent Attacks
We first consider the coefficient a1 in the attack model,
which controls the effect of the proportion of co-ethnics
on the likelihood of attacks. Figure 3 plots the density of
a1 in H0 (solid line) against the full sampling range (grey
line), which we drew from a uniform distribution.
Recall that a negative estimate for a1 would be consistent with the “maximum impact” mechanism of violence:
Attacks become more likely the higher the proportion of
the other group in order to maximize damage to the outgroup. However, the plot shows that across all plausible
parameter vectors, a1 is almost consistently positive. This
leads to two conclusions: First, a violence generating
mechanism that depends on the local ethnic configuration is a necessary part of the model. In other words, we
can show that a random allocation of insurgent attacks,
independently of the local ethnic configuration, would
not account for the patterns we observe in Baghdad.
Moreover, the positive coefficient provides support for an
“ethnic cleansing” type of violence, where higher proportions of co-ethnics encourage insurgents to attack small
minorities of the other group. This is consistent with the
argument that insurgents need a local support base in
order to successful carry out attacks and that attacks are
motivated by the desire to create ethnically homogenous
enclaves.
11
The model dynamics start at time step 1.
60
Violence and Ethnic Segregation
0.00
0.05
Density
0.10
0.15
against one’s ethnic group are positively related to migration, but only in the locality where attacks occur. In
short, as we would expect, violence in a locality causes
members of the affected ethnic group to move to other
locations.
−10
−5
0
5
10
α1 (proportion co−ethnics)
FIG 3. Density of a1 in H0 (solid line). The full sampling range in Θ
is shown as a grey line
0.04
Density
0.03
0.00
0.00
0.01
0.02
0.05
Density
0.10
0.05
0.06
0.07
0.15
Experienced Violence and Migration
Next, we consider the effect of violence on migration, as
captured by the coefficients b1 and b2 in the migration
model. b1 determines the effect of violence in a civilian’s
neighborhood on her decision to migrate, whereas b2
captures the indirect effect of violence occurring in the
proximity of the civilian (the spatial lag of violence). The
densities of the two coefficients are shown in Figure 4.
For b1 , we see a result similar to the previous one: The
density of the parameter is clearly positive, as we would
expect. We can conclude that fear created by insurgent
violence has a strong impact on the settlement pattern
changes we see in Baghdad during the course of our
study period. The right panel shows that, counter to our
expectations, there does not seem to be a signaling effect
of violence across neighborhoods. The spatially lagged
violence does not seem to affect an individual’s migration
decision in a clearly discernible way, since the density of
the b2 in H0 is virtually indistinguishable from the sampling density and can take both negative and positive values. In sum, our models produce a better fit when attacks
Violence and Segregation over Time
One of our principal claims is that there is a reciprocal
relationship between violence and settlement patterns.
Violence should induce migration, but ethnic homogeneity should reduce levels of violence, even in the absence
of policing. In order to show how our key outputs of
interest–the level of violence and segregation–evolve over
time, we recorded these outputs at every time step in the
533 model runs with parameters in H0 . Averages over
these runs are plotted in Figure 5.
Clearly, increased segregation is generally related to
lower levels of violence, which is also what we observed in
Baghdad. However, what is the effect of local ethnic mixing on violence, i.e. how does the composition of neighborhood affect its risk of ethnic violence? Our results
above suggest the following relationship: As mixed ethnic
neighborhoods move towards homogeneity, attacks
against the minority increase (positive effect of proportion co-ethnics). At the same time, however, the outflow
of minority members reduces the number of potential
targets, removing the risk of violence in homogenous
neighborhoods. Consequently, there should be a curvilinear relationship between segregation and violence, with
intermediate levels of segregation giving rise to many
attacks, but a decline at high levels of segregation. A
quick additional test confirms this. Table 1 shows the
results of a regression of the (logged) number of attacks
on the proportion of homogenous neighborhoods and its
squared term (including dummies for each model run),
which confirms the inverted U-shaped relationship.
What can we conclude from these results? First, we are
able to show that violent attacks are consistent with an
ethnic-cleansing mechanism: Attacks are more likely to
occur in areas where there are small but significant
minorities. Second, we show that violence increases the
level of ethnic segregation over time. Significantly, we do
not need to assume ethnophilia to generate this result.
Ethnic segregation need not be driven by antipathy
toward out-groups; the search for safety is all that is
required to produce unmixing. Finally, we show that
−10
−5
0
5
β1 (experienced violence)
(A) Experienced violence
10
−10
−5
0
5
10
β2 (violence, spatial lag)
(B) Experienced violence (spatial lag)
FIG 4. Density of b1 (left panel) and b2 (right panel) in H0 (solid lines)
61
0.6
0.0
7.5
0.2
0.4
Proportion of homogenous units
9.0
8.5
8.0
Number of attacks
9.5
0.8
10.0
1.0
Nils B. Weidmann and Idean Salehyan
0
100
200
300
400
500
Time
0
100
200
300
400
500
Time
(A) Violence over time
(B) Segregation over time
FIG 5. Evolution of violence (left) and segregation (right) over time, averaged over 533 models runs with parameters from H0
Implementing Policing
TABLE 1. Regression of violence on segregation
Model 1
Prop. homogeneous
Prop. homogeneous (squared)
(Intercept)
3.73* (0.03)
3.21* (0.03)
0.45* (0.02)
Number of observations: 266,500. Adjusted R 2 : 0.83. Coefficients for the
model run dummies not shown. Standard errors in parentheses.
significance at p < 0.05.
ethnic segregation–by itself–can reduce the level of violence in a system, even if it is not normatively desirable.
Under conditions of anarchy, violence declines as neighborhoods become more homogenous–a finding in support of arguments that the troop surge was irrelevant for
the drop in observed violence in Baghdad (Agnew et al.
2008). Thus, there is a reciprocal relationship between
settlement patterns and violence: Initial levels of ethnic
mixing influence where attacks are likely to occur; these
attacks produce migration, and in turn, segregation
reduces the level of violence. In the next section, we turn
to policing and its ability to limit insurgent attacks.
Policing to Reduce Ethnic Violence
In this section, we provide a more nuanced perspective on
alternative solutions to ethnic violence, in particular those
that involve policing by outside actors (e.g. the state).
Based on our above results, we cannot dismiss the effect
of policing altogether; all we can say is that segregation is
one way to reduce violence, and could have been sufficient
to bring down violence in Baghdad. To what extent can
policing be successful in limiting violence? Can enhanced
policing effots be effective relatively late in an insurgency,
or is it better to have a robust force in place early on, as
argued by General Shinseki? We address these questions
in two steps. First, we analyze an “early on” implementation of policing, with policing starting at time step 1. In
these simulations, we systematically vary the level of policing success to assess its effect on the reduction of violence
and ethnic segregation. Second, we perform a set of simulation runs to examine more closely the effect of timing,
i.e. the time when counterinsurgent efforts start. This
helps us to disentangle whether the size of the policing
force, or the timing of these efforts matter more.
As argued above, an alternative way of violence reduction
is by means of policing. By adding a simple policing
mechanism to our model, we can “re-run” history and
assess different counterfactual worlds. Policing is modeled
using a “success rate” of counterinsurgency, which
accounts for the probability that an insurgent is removed
from the scene by capture or killing, after having carried
out an attack. Note that in this implementation, agents
can only be captured once they have attacked successfully
and thus have visibly identified themselves as insurgents.
In the present version of the model, we remain agnostic
about possible variation in this success rate across locations, and assume a constant probability of capture. After
an insurgent has carried out an attack, we take a random
draw from a Bernoulli distribution with success rate τ. If
successful, the insurgent will be removed from the model.
In addition, we can “switch on” policing at a particular
time during the simulation. Furthermore, we assume that
punished attacks do not generate fear in the targeted
population in the same way as regular attacks do. Therefore, only the number of unpunished attacks enters the
agents’ migration decision.12
We apply our counterfactual experiments—the variations in policing success and timing—to the parameter
estimates obtained above (i.e. the parameter vectors in
H0 , with a 0.02 proportion of insurgents). We initialize
the model using the ethnic map of 2003, which reflects
the distribution of groups in Baghdad roughly at the
beginning of the invasion. This map shows that Baghdad
is much less segregated than in 2006, with only about
30% of the neighborhoods coded as homogenous. Based
on this setup, we systematically vary the success rate of
policing (experiment 1) and the onset of policing (experiment 2). For each value of these experimental variations,
we run the model on all 533 parameter vectors in H0 ,
again limiting each run to 500 time steps. We collect
information about the level of violence (measured as
the total number of attacks in a run) and segregation
(measured as the proportion of homogenous units). The
results shown below are averages over the 100 parameters
combinations tested for each experimental variation.
12
Our results do not depend crucially on this assumption; see online
appendix.
62
Violence and Ethnic Segregation
these results emphasize even more the importance of a
timely response. Thus, these results may be seen as a vindication of the Shinseki plan: early implementation of
effective counterinsurgency can have a very large effect,
while late implementation is much less successful.
Conclusion
In this paper, we examined the relationship between violence and ethnic segregation using an agent-based model
informed by actual data from Baghdad. While we are not
able to re-run history and draw definitive conclusions
about what actually happened in Iraq, our results do contribute to our understanding of ethnic conflict more generally. First, we show that ethnic settlement patterns
influence where violent attacks are likely to occur. Violence is most likely when there are small but significant
minorities in a locality as insurgents attempt to ethnically
cleanse neighborhoods. Second, the desire for safety is
enough to generate a considerable degree of ethnic segregation. Although people may desire to live in neighborhoods dominated by their co-ethnics, this is not necessary
to produce ethnic unmixing in the context of escalating
violence. All that is needed is for individuals to move to
areas that are relatively safer. Third, we show that as
neighborhoods become more segregated, the level of violence eventually declines. While partition into mutually
exclusive ethnic enclaves may not be normatively desirable, it can help to reduce bloodshed in the short run.
Thus, there is a reciprocal relationship between ethnic
segregation patterns and violence. While we cannot provide the final answer as to what happened in Baghdad,
our results indicate that one plausible explanation for
why violence declined as it did was due to the fact that by
the time of the troop surge, ethnic segregation was nearly
complete.
Our modeling exercise also informs current debates
about the effectiveness of counterinsurgency operations.
We show that even small increases in policing effectiveness can dramatically mitigate the level of violence, but
only if policing is implemented very early on. Delayed
implementation of a robust counterinsurgency strategy
may help to reduce violence somewhat, but it is not
nearly as effective as efforts early on. This finding has
important implications for policy discussions. While
increasing the number of troops in Iraq and Afghanistan
0.6
0.5
0.4
0.3
0.2
0.0
0.1
Proportion of homogeneous units
3000
2000
1000
0
Total number of attacks
4000
Our first counterfactual experiment tests the effect of
increased policing success on violence and segregation,
and Figure 6 shows the results. We vary the policing success rate τ from 0 to 1 in increments of 0.1; policing is
implemented from the beginning of the simulation. The
left panel shows that even small levels of policing are
highly effective in reducing violence: Even a low success
rate of 0.1 reduces violence by more than 75% compared
to no policing. The effects of policing on segregation are
not as strong. There is a continuous decline in segregation as policing increases, but with limited success. Even
implausibly high policing success rates of 0.5 and above
cannot prevent a segregation up to a certain limit (recall
that the initial proportion of homogenous units is 0.3).
The reason is that for policing to be effective and insurgents to be captured, these insurgents have to attack first.
This unavoidable level of violence may be sufficient to
trigger the ethnic unmixing we see in this experiment.
Nonetheless, we can conclude from this experiment that
policing can have a large impact on reducing violence
levels, even at relatively low levels of counterinsurgent
effectiveness.
In our first experiment described above, the policing
mechanism is active from the first time step. In the next
experiment, we examine the timing of policing, varying
the onset of policing from time step 0 to 500 in steps of
50. We do this for two different policing success rates:
low (τ = 0.1), medium (τ = 0.3) and high (τ = 0.5). Figure 7 shows the results for low (solid lines), medium
(dashed) and high (dotted lines) levels of success. As we
expected, the later the policing efforts start, the smaller
the reduction in violence and segregation. Again, similar
to the previous experiment, the effect on segregation is
limited due to the migration occurring even with low levels of violence. Both plots, however, draw our attention
to the importance of the timing, especially the first 100
time steps. In order to compensate for a delay of 100
time steps in the onset of our policing efforts, we would
have to roughly triple policing efforts to achieve a comparable reduction in violence: we achieve the same reduction in violence if we implement policing with a low (0.1)
success rate right from the beginning, or with a medium
(0.3) success rate after 100 time steps. The effect on segregation reduction is similar. Together with our results
from the previous experiments, which showed that even
small success rates can mitigate violence considerably,
0.0
0.2
0.4
0.6
Policing success rate
(A) Violence
0.8
1.0
0.0
0.2
0.4
0.6
0.8
Policing success rate
(B) Segregation
FIG 6. Effect of policing on the reduction of violence (left) and segregation (right)
1.0
63
0
100
200
300
400
500
Policing onset (time step)
(A) Violence
0.60
0.55
0.50
0.45
0.40
Proportion of homogeneous units
Policing
low (0.1)
medium (0.3)
high (0.5)
0.35
low (0.1)
medium (0.3)
high (0.5)
0.30
3000
2000
1000
Policing
0
Total number of attacks
4000
0.65
Nils B. Weidmann and Idean Salehyan
0
100
200
300
400
500
Policing onset (time step)
(B) Segregation
FIG 7. Effect of the onset of policing on the reduction of violence (left) and segregation (right)
may have some utility in the context of a raging insurgency, resources would have been better used if sufficient
forces were in place from the start. In the future, policy
makers must exercise extreme caution before engaging
in military occupations, and do so only if they are willing
to devote overwhelming resources to contain an insurgency.
Finally, in this paper we present an approach to agentbased modeling that is grounded in empirical data. We
use actual data from Baghdad to initialize our model,
and we attempt to select those parameters that best fit
the empirical record. While abstract modeling exercises
can be extremely useful in developing theories and contributing to scholarly debate, validating these models with
empirical data can help to demonstrate the utility of a
particular tool. No computational model can credibly
claim to perfectly recreate history, but such tools can significantly aid in our understanding of complex social
phenomena.
References
Agnew, John, Thomas W. Gillespie, Jorge Gonzalez, and Brian Min.
(2008) Baghdad Nights: Evaluating the US Military “Surge” Using
Nighttime Light Signatures. Environment and Planning A 40: 2285–
2295.
Axelrod, Robert. (1984) The Evolution of Cooperation. New York: Basic
Books.
Bennett, D. Scott. (2008) Governments, Civilians, and the Evolution of
Insurgency: Modeling the Early Dynamics of Insurgencies. Journal of
Artificial Societies and Social Simulation 11 (4).
Berman, Eli, Jacob N. Shapiro, and Joseph H. Felter. (2011) Can
Hearts and Minds Be Bought? The Economics of Counterinsurgency
in Iraq. Journal of Political Economy 119 (4): 766–819.
Bhavnani, Ravi. (2006) Ethnic Norms and Interethnic Violence:
Accounting for Mass Participation in the Rwandan Genocide.
Journal of Peace Research 43 (6): 651–669.
Bright, Arthur. (2007) Baghdad’s Sunni/Shiite Security Wall.
Christian Science Monitor, April 20.
Cederman, Lars-Erik. (1997) Emergent Actors in World Politics: How States
and Nations Develop and Dissolve. Princeton, NJ: Princeton University
Press.
Cederman, Lars-Erik. (2001) Agent-Based Modeling in Political
Science. The Political Methodologist 10: 16–22.
Cederman, Lars-Erik. (2003) Modeling the Size of Wars: From Billiard
Balls to Sandpiles. American Political Science Review 97 (1): 135–150.
Cederman, Lars-Erik. (2008) Articulating the Geo-Cultural Logic of
Nationalist Insurgency. In Order, Conflict, and Violence, edited by
Stathis Kalyvas, Ian Shapiro and Tarek Masoud. Cambridge, UK:
Cambridge University Press.
Chapman, Thomas, and Philip G. Roeder. (2007) Partition as a
Solution to Wars of Nationalism: The Importance of Institutions.
American Political Science Review 101 (4): 677–691.
Cioffi-Revilla, Claudio, and Mark Rouleau. (2010) MASON
RebeLand: An Agent-based Model of Politics, Environment, and
Insurgency. International Studies Review 12 (1): 31–52.
Coleman, James S. (1990) Foundations of Social Theory. Cambridge, MA:
The Belknap Press of Harvard University Press.
Collier, Paul, and Anke Hoeffler. (2004) Greed and Grievance in
Civil War. Oxford Economic Papers 56 (4): 563–595.
Dean, Jeffrey S., George J. Gumerman, Joshua M. Epstein, Robert
Axtell, Alan C. Swedlund, Miles T. Parker, and Stephen
McCarroll. (2006) Understanding Anasazi Culture Change
through Agent-based Modeling. In Generative Social Science: Studies in
Agent-based Computational Modeling, edited by Joshua M. Epstein.
Princeton, NJ: Princeton University Press.
de Marchi, Scott. (2005) Computational and Mathematical Modeling in the
Social Sciences. New York: Cambridge University Press.
Enterline, Andrew J., and Michael Greig. (2007) Surge, Escalate,
Withdraw, and Shinseki: Forecasting and Retro-Casting American
Force Strategies and Insurgency in Iraq. International Studies
Perspectives 8 (3): 245–252.
Epstein, Joshua M. (2002) Modeling Civil Violence: An Agent-based
Computational Approach. Proceedings of the National Academy of
Sciences USA 99 (3): 7243–7250.
Epstein, Joshua M., and Robert Axtell. (1996) Growing Artificial
Societies. Cambridge, MA: MIT Press.
Fearon, James D., and David D. Laitin. (2003) Ethnicity, Insurgency
and Civil War. American Political Science Review 97 (1): 75–90.
Geller, Armando., Seyed M. M. Rizi, and Maciej M. Łatek. (2011)
How Corruption Blunts Counternarcotic Policies in Afghanistan: A
Multiagent Investigation. In Social Computing, Behavioral-Cultural
Modeling and Prediction, 2011 Conference. Lecture Notes in Computer
Science 6489, edited by John Salerno, Shanchieh Jay Yang, Dana
Nau, and Sun-Ki Chai. New York: Springer.
Geller, Armando, and Shah Jamal Alam. (2010) A Socio-Political and
-Cultural Model of the War in Afghanistan. International Studies
Review 12 (1): 8–30.
Hammond, Ross, and Robert Axelrod. (2006) The Evolution of
Ethnocentrism. Journal of Conflict Resolution 50 (6): 926–936.
Hendrix, Cullen. (2010) Measuring State Capacity: Theoretical and
Empirical Implications for the Study of Civil Conflict. Journal of
Peace Research 47 (3): 273–285.
Herbst, Jeffrey. (2000) States and Power in Africa: Comparative Lessons in
Authority and Control. Princeton, NJ: Princeton University Press.
64
Violence and Ethnic Segregation
Herbst, Jeffrey. (2004) African Militaries and Rebellion: The Political
Economy of Threat and Combat Effectiveness. Journal of Peace
Research 41 (3): 357–369.
Joseph, Edward, and Michael E. O’Hanlon. (2007) The Case for Soft
Partition of Iraq. Saban Center for Middle East Policy Analysis
Papers, 12 (June).
Kalyvas, Stathis N. (2006) The Logic of Violence in Civil War. New York:
Cambridge University Press.
Kaufmann, Chaim. (1996) Possible and Impossible Solutions to Ethnic
Civil Wars. International Security 20 (4): 136–175.
Lake, David. (2002) Rational Extremism: Understanding Terrorism in
the Twenty-First Century. Dialogue-IO 1 (15): 29.
Londono, Ernesto. (2007) In Baghdad, a Flimsy Outpost. Washington
Post, March 22.
Lustick, Ian S., Dan Miodownik, and Roy J. Eidelson. (2004)
Secessionism in Multicultural States: Does Sharing Power
Prevent or Encourage it?American Political Science Review 98 (2):
209–229.
Lyall, Jason, and Isaiah Wilson. (2009) Rage Against the Machines:
Explaining Outcomes in Counterinsurgency Wars. International
Organization 63 (1): 67–106.
Matthews, Dylan. (2008) How Important Was the Surge? The American
Prospect Webzine. Monday, July 28, 2008.
McCain, John, and Joe Lieberman. (2008) The Surge Worked. Wall
Street Journal. Opinion. January 10, 2008.
Mueller, John. (2000) The Banality of “Ethnic War.” International
Security 25 (1): 42–70.
Muller, Edward, and Mitchell Seligson. (1987) Inequality and
Insurgency. American Political Science Review 81 (2): 425–451.
Odierno, Raymond T. (2008) The Surge in Iraq: One Year
Later. Lecture on National Security and Defense, The Heritage
Foundation, March 13. Available at http://www.heritage.org/
Research/Lecture/The-Surge-in-Iraq-One-Year-Later. (Accessed April
1, 2010.)
Olzak, Susan, Susanne Shanahan, and Elizabeth McEneaney. (1996)
Poverty, Segregation, and Race Riots: 1960–1993. American
Sociological Review 61 (4): 590–613.
Petraeus, David H. (2007) Report to Congress on the Situation in Iraq.
September 10–11, 2007.
Posen, Barry R. (1993) The Security Dilemma and Ethnic Conflict. In
Ethnic Conflict and International Security, edited by Michael E. Brown.
Princeton, NJ: Princeton University Press.
Raghavan, Sudarsan. (2007) Marriages Between Sects Come Under
Siege in Iraq. Washington Post, March 4.
Ricks, Thomas. (2009) The Gamble: General David Petraeus and the
American Military Adventure in Iraq, 2006–2008. New York: Penguin.
Rose, William. (2000) The Security Dilemma and Ethnic Conflict: Some
New Hypotheses. Security Studies 9 (4): 1–51.
Sambanis, Nicholas. (2000) Partition as a Solution to Ethnic War. World
Politics 52: 437–483.
Schelling, Thomas C. (1971) Dynamic Models of Segregation. Journal
of Mathematical Sociology 1: 143–186.
Siegel, David A. (2009) Social Networks and Collective Action. American
Journal of Political Science 53 (1): 122–138.
Telhami, Shibley. (2005) Rush To Stabilize May Backfire in Polarized
Iraq. San Jose Mercury News, October 30.
Tilly, Charles. (1978) From Mobilization to Revolution. Reading:
Addison-Wesley.
Urdal, Henrik. (2006) A Clash of Generations? Youth Bulges and
Political Violence. International Studies Quarterly 50 (3): 607–629.
Wilkinson, Steven. (2004) Votes and Violence. Cambridge, UK:
Cambridge University Press.
Wong, Edward, and David Cloud. (2007) US Erects Baghdad Wall to
Keep Sects Apart. New York Times, April 21.